CoVar extracts intelligence from unstructured sensing data with our machine learning and signal processing technologies.
CoVar has world-class expertise in development and implementation of cutting-edge software systems with a focus on processing sensor data, automated extraction of information, and automated decision-making. Adaptive, dynamic software act as the brain for a sensor system. Our principals and staff have pioneered advanced techniques in image processing, pattern recognition, and classification algorithms. By leveraging cutting-edge statistics and new models that are being developed daily, we develop software that is physics-based, statistically robust, adaptive, and reliable.
Machine learning algorithms enable computers to learn and adapt just as a human being does. CoVar principals are published experts in the field of machine learning software. Machine learning algorithms, founded on the latest statistical and computer engineering principles, are designed to allow the software to adapt to the particular properties of the data being processed, whether real-time sensor data, metadata tags, or simulated data. By codifying this adaptivity, data can be organized, meaningful information can be extracted, and automated decisions can be optimally made with little or no human intervention. CoVar has developed and fielded a variety of machine learning algorithms with varying degrees of autonomy and adaptivity, depending on the particular requirements of the application.
Signal and Image Processing
Sensors provide data, whether simple time-based measurements (e.g. pressure, flow) or less structured, multi-dimensional measurements (e.g. video, IR, spectral). Effective signal and image processing optimally extracts the most meaningful information (signals) from the distracting or confusing components of that data (noise). From pre-processing and feature-extraction to learning classifiers, these algorithms transform raw data into cleaner, more information-rich data. The development of effective data-to-answers signal and image processing algorithms is as much an art as it is a science. By combining our knowledge of time-tested methods, the latest breaking methods, and our physics-based understanding of the sensing domain, we develop algorithms that work, work reliably, and generalize across datasets collected under varying conditions.
Often sensing problems are so difficult or the accuracy requirements are so high (e.g. standoff landmine detection) that a single type of sensor, even with the most sophisticated algorithms, simply does not provide enough information to make reliable decisions. In these cases, multiple sensor modalities might be combined to provide orthogonal information about the sensing domain. For example, landmine detection might be accomplished with standoff ground penetrating radar together with visual clues about disturbed earth and the presence of ground clutter. Combining the data from these disparate sensors in ways that optimally improve the system performance is nontrivial. CoVar principals are pioneering approaches using novel probabilistic models for multi-modal sensor fusion that maximize the information to be gained from orthogonal sensing.
CoVar principals have spent the last two decades innovating, developing, and fielding automatic target recognition algorithms and signal processing for sub-surface sensing technologies. We have developed algorithms for IED and landmine detection leveraging ground penetrating radar sensors, seismic and acoustic sensors, and infrared modalities. We have provided enhanced intelligent software for municipal and commercial infrastructure inspection products, and we have developed statistics- and physics-based predictive techniques for oil and gas drilling applications.
CoVar scientists are recognized subject matter experts in classification with spectroscopic data, including laser-induced breakdown spectroscopy (LIBS), RAMAN, and more. Our proprietary software and data analytics packages work with spectroscopy technologies to make statistically robust inferences about underlying material properties. Our techniques are especially relevant in noisy environments or environments with high clutter signatures. CoVar has developed and sells a powerful classification software package for the industry, SpectraLearn. This easy-to-use, yet sophisticated toolkit puts powerful machine learning capabilities at the fingertips of the researcher or developer. For more information about SpectraLearn, click here.
As the principal developers of the “Pattern Recognition Toolbox”, an open-source and widely used MATLAB plug-in, we have in-depth and unparalleled understanding of how machine learning techniques and classification can be applied to a wide variety of challenging data analysis and sensing problems. Reliable detection in noisy environments, multi-dimensional anomaly classification, clutter tracking and suppression—really, any application requiring real-time robust extraction of information from a complicated, large, and noisy set of data—we can help. For more information on the PRT click here.